Add paper link, GitHub link and task category to dataset card

README.md

@@ -52,4 +52,33 @@ configs:
   data_files:
   - split: test
     path: data/test-*
+task_categories:
+- video-text-to-text
 ---
+
+# AVQA-Hard
+
+AVQA-Hard is a curated subset of the AVQA dataset, introduced in the paper "[Do Modern Video-LLMs Need to Listen? A Benchmark Audit and Scalable Remedy](https://huggingface.co/papers/2509.17901)".
+
+This dataset filters out items that are solvable from a single frame (visual shortcuts) to genuinely probe the audio-visual reasoning capabilities of Video-LLMs. The authors found that many standard benchmarks do not adequately penalize models that ignore audio, whereas this "Hard" split makes audio cues decisive for correct predictions.
+
+## Project Resources
+- **Paper:** [Do Modern Video-LLMs Need to Listen? A Benchmark Audit and Scalable Remedy](https://huggingface.co/papers/2509.17901)
+- **GitHub:** [https://github.com/naver-ai/LLaVA-AV-SSM](https://github.com/naver-ai/LLaVA-AV-SSM)
+
+## Usage
+These files are formatted for compatibility with [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval).
+
+## Citation
+If you find this work or dataset useful, please cite:
+```bibtex
+@misc{kim2025doesaudiomattermodern,
+  title         = {Does Audio Matter for Modern Video-LLMs and Their Benchmarks?},
+  author        = {Geewook Kim and Minjoon Seo},
+  year          = {2025},
+  eprint        = {2509.17901},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.CV},
+  url           = {https://arxiv.org/abs/2509.17901}
+}
+```